Towards On-line Learning Agents for Autonomous Navigation
Doktorsavhandling, 1999

The design of a mechatronic agent capable of navigating autonomously in a changing and perhaps previously unfamiliar environment is a very challenging issue. This thesis addresses this issue from both functional and system perspectives. Functions such as spatial representation, localization, path-finding and collision avoidance are essential to autonomous agentn. Four types of learning related to these functions have been identified as important: sensory information categorization and classification, the learning of stimulus-response mapping, the learning of spatial representation and the coding and adaptation of the travel experience with regard to specific tasks. It is argued that, in order to achieve a high degree of autonomy at the system level, it is essential to implement each of these navigational functions with a highly autonomous learning technique. An analysis of several representative artificial neural network (ANN) algorithms for their degrees of autonomy and computational characteristics indicates that none of the learning techniques analyzed is alone sufficient in terms of spatial learning. It is shown that biology can be inspirational in finding a possibly better, or perhaps more complete, solution to the learning of spatial representation than previous engineering or ANN based approaches. In particular, data on the biological head direction system have inspired the generation of a computational model which is shown to be able to use learned environmental features to correct the directional error accumulated by dead-reckoning in a simulated mobile robot. Furthermore, using a hippocampal place learning system in biological systems as an inspiration, a network model of dynamic cell structure is suggested. It allows an autonomous agent to perform tasks such as environmental mapping, localization and path-finding. In this model, a focus mechanism is included to help minimize computation needs by directing the adaptation of the network and the pathfinding. The thesis also discusses various approaches toward achieving a high degree of autonomy at the system level. It is also shown that a feed forward gating mechanism can be combined into a layered design framework to accommodate the interaction between various navigational functions having high degrees of autonomy.

path-finding

head direction

artificial neural network

spatial representation

on-line learning

navigation

localization

autonomous mobile robot

obstacle avoidance

place learning

Författare

Guang Li

Institutionen för datorteknik

Ämneskategorier

Data- och informationsvetenskap

ISBN

91-7197-824-0

Doktorsavhandlingar vid Chalmers tekniska högskola. Ny serie: 1523

Technical report - School of Electrical and Computer Engineering, Chalmers University of Technology, Göteborg, Sweden: 371